Abstract:
Cassava is a nutty-flavored long bulbaceous starchy root vegetable. It is the
principal source of calories and carbs for many people around the world,
especially in southern Africa. Cassava production is most common in South
Africa because it can survive in harsh environments. Sometimes the cassava
crop gets affected by leaf disease, which infects its overall production and
reduces the farmers’ income. And manual leaf disease detection may not
obtain proper accuracy. In order to detect cassava leaf diseases, the current
studies face various challenges such as poor accuracy, low detection rate, and
high processing time. In our research, we have used supervised contrastive
learning to detect four diseases of the cassava leaf and identify healthy leaves,
from which we got tremendous results. We also used data augmentation
modules, encoder networks, and projection networks to perform certain tasks
such as network training, embedding similar classes nearby and other classes
away, image labeling, and so on. From our study, we have achieved accuracy,
precision, recall, and f1score of 88%, 78%, 79%, and 79%, respectively using
the supervised contrastive model.
